Performance Considerations When Using a Dedicated Ray Traversal Engine

Abstract

In the recent years we have witnessed massive boost to hardware graphics accelerators (graphics cards), not only in the raw performance, but also in their programmability, introducing the concept of GPGPU. However, despite this, the current architectures still favor feed-forward algorithms over recursive ones. While shading is, in this sense, a feed-forward algorithm, ray tracing, and specifically ray traversal, is a recursive rather than feed-forward algorithm. Adding a dedicated hardware ray traversal engine should therefore prove to be an interesting option. Also, with dedicated hardware, we can perform many optimizations on arithmetic units due to their fixed interaction. This can reduce the area well below a simple sum of areas of the individual units. In this paper we offer for consideration analysis of memory requirements for combination of a dedicated hardware ray traversal and intersection engine with highly parallel general purpose processor used for shading. We show results and requirements of such a combination on scenes of moderate complexity, with regard to speed, bandwidth and latency.